Timing samples measure the selected optimizations in fixed-shape builds. Installation and general-dimension correctness checks are recorded separately.
The JSON dataset includes every per-run summary used in the current tables, including both orders. It contains no captured tensors, GPU code or personal machine paths. Source-report SHA-256 values document the export provenance; the full research reports remain private.
Run with Python 3.9 or later (standard library only):
python tools/report_results.pyThis recomputes the displayed tables from the exported measurements. It does not launch the game or reproduce the neural network without its missing dependencies.
RTX 4080, driver 616.56, model1/sRGB with fixed synthetic inputs. 100 warmup frames, 400 measured frames per run. Independent-process rows average each run's P50; interleaved rows average the two runs' arm means. For each displayed row:
time reduction = 100 * (1 - mean(optimized) / mean(community)).
ABBA means community, optimized, optimized, community; BAAB reverses those labels. Both orders are retained to reduce ordering bias. Interleaved arms share history, so separate independent output checks are required. The two 4K interleaved runs have substantial absolute-time drift; neither is discarded. Independent 4K timing supports 8.16%, while interleaved timing gives 10.30%. Do not report a guaranteed 10% reduction from these observations.
1440p independent-process timing is missing, not zero. The previous 4K 10.20% summary used a median of within-run group percentages; 10.30% uses the current table's aggregate-ratio formula and is not a newly obtained speedup.
Cyberpunk 2077 2.31's built-in benchmark, High raster, native DLAA and SDR. Disable RT, PT, RR, frame generation, dynamic resolution, Reflex, VSync and the frame cap. Use the same settings for original game, community and optimized within a cohort. Original game has no experimental NR/ReShade stack. Each role has two runs. The three resolutions are separate complete cohorts, not one continuous session.
FPS gain = 100 * (mean(optimized FPS) / mean(community FPS) - 1).
added frame ms = mean(1000 / enabled FPS per run) - mean(1000 / original FPS per run).
Do not replace mean(1000/FPS) with 1000/mean(FPS) or interpret added frame cost
as a timer around the neural network. The delta includes game insertion, waits and
resource competition. GPU clocks were not locked, and two repeats cannot establish
robust confidence intervals. The latest 1080p paired FPS gains are 0.42%/2.33%;
the earlier complete cohort's 0.40% remains an effective warning about variability.
A failed original-game 1440p attempt was excluded before a complete 1440p cohort was rerun. Failed runs and old anomalous output controls are not performance samples. Package identities are checked against accepted game routing evidence. Completed short benchmarks are not proof of long-session stability or universal visual equivalence.
Tim Schiesser's TechSpot test published on September 4, 2026 reports 7.8 ms at 1440p and 17.0 ms at 4K in its “Measured DLSS 5 Render Times” table. Its game is NBA 2K27 with official DLSS5 integration. It estimates cost from on/off frame rates and adjusts some GPU settings to DLAA to avoid bottlenecks. Its render-time table has no 1080p row.
numerical difference from reference = 100 * (1 - our added ms / external ms).
This is 10.66% at 1440p and 9.32% at 4K. It is a comparison of published estimates under different conditions, not a controlled GPU ranking, FPS gain, or pure NR inference speedup. A matched-game/platform comparison remains on the roadmap.